tom/kyutai-moshi
1
0
Fork 0
Code Issues Pull requests Packages 1 Activity Actions
686 commits 14 branches 24 tags 5.8 MiB
Commit graph

686 commits

This branch
This branch
All branches
Author SHA1 Message Date
Tom Foster
2c902a1369 Forgejo build
All checks were successful
Build and Push Docker Image / build (push) Successful in 51m41s
Details
2025-07-14 17:20:44 +01:00
Laurent Mazare
97868dbbb0
Moshi mlx.0.2.11 (#333) 
* Fix the local-web MLX example + port to sphn 0.2.

* Formatting fixes.

* Use sphn 0.2.

* Fix.

* Bump the moshi mlx dependency.
 2025-07-08 19:22:18 +02:00
Laurent Mazare
ec871a6c0e
Fix the local-web MLX example + port to sphn 0.2. (#332) 
* Fix the local-web MLX example + port to sphn 0.2.

* Formatting fixes.

* Use sphn 0.2.

* Fix.
 2025-07-08 18:26:33 +02:00
Laurent Mazare
459add59e6
Extra-heads support for the MLX implementation. (#331) 
* Extra-heads support for the MLX implementation.

* Tweaks.
 2025-07-08 16:01:16 +02:00
Laurent Mazare
7b5f6505ca
Sketch the 0.2.9 release. (#330) 2025-07-08 10:51:01 +02:00
Laurent Mazare
9837ca328d
Allow torch 2.7. (#329) 2025-07-07 14:30:07 +02:00
Laurent Mazare
2efed9df3d
Extra-heads support in the pytorch implementation. (#327) 
* Extra-heads support in the pytorch implementation.

* Run the extra-heads.

* Formatting.

* Fix + alpha1.
 2025-07-07 12:34:35 +02:00
Laurent Mazare
faae3a5225
Bump the moshi dependency used on the rust side to 0.2.8. (#326) 2025-07-05 10:30:43 +02:00
Alexandre Défossez
3b1c248b80 v0.2.8 release 2025-07-04 15:03:27 +02:00
laurent
fbc91dbfa6 Tweak the (unused) python entry points names. 2025-07-04 08:36:07 +02:00
Alexandre Défossez
96b0ebce50 Merge branch 'main' of github.com:kyutai-labs/moshi 2025-07-03 13:49:50 +02:00
laurent
ee4fd743c4 Upgrade the rust version. 2025-07-03 13:43:15 +02:00
laurent
a40c5612ad Update the moshi package in the rust dependency. 2025-07-03 13:36:08 +02:00
Alexandre Défossez
8b8a88fc4f v0.2.7 moshi release 2025-07-03 13:24:50 +02:00
laurent
b2ec69e925 Update to the latest rust version. 2025-07-03 12:28:28 +02:00
Laurent Mazare
c84a61ad70
Bump the mlx version. (#325) 2025-07-03 09:39:20 +02:00
Laurent Mazare
4033ef60e2
Add the on_frame callback to moshi mlx too. (#324) 2025-07-03 09:06:18 +02:00
Laurent Mazare
c8f29b8b28
Add an on-frame callback to moshi pth tts. (#323) 2025-07-03 08:15:11 +02:00
Laurent Mazare
474daba455
Add some hooks to handle streaming. (#322) 2025-07-02 16:08:49 +02:00
Alexandre Défossez
192b9d82d4
plop (#319) 2025-07-02 15:08:16 +02:00
Laurent Mazare
7b684410aa
Bump the mlx version. (#321) 2025-07-02 14:52:06 +02:00
Laurent Mazare
cfd3c56b30
Move the mlx pytorch loader. (#320) 2025-07-02 14:40:24 +02:00
Laurent Mazare
1fe6aa377f
Add a quatization option for the mlx tts. (#318) 2025-07-02 12:53:20 +02:00
Laurent Mazare
af077f39bc
Another mimi mlx fix. (#317) 2025-07-02 11:48:09 +02:00
Eugene Kharitonov
505c864771
Adding on_text_logits_hook (#314) 
which is supposed to be used for various controlled generation
tricks -- effectively allows to modify the sampling probabilities and or
sampling space.

Co-authored-by: Eugene <eugene@kyutai.org>
 2025-07-02 11:17:25 +02:00
Laurent Mazare
84d49e9ed9
Fix mimi for the recent versions of MLX. (#316) 2025-07-02 10:54:58 +02:00
Laurent Mazare
56ddf28625
MLX implementation for the TTS (#313) 
* MLX implementation of the TTS.

* More MLX TTS work.

* Add the required conditioners.

* Add more conditioners.

* More work for the MLX TTS.

* Fix the shape issues.

* Support for low rank embeddings.

* Cosmetic things.

* Voice loading.

* Tweaks.

* Load the weights.

* Start adding the cross-attention.

* Add some cross-attention.

* Pre-commit fix.

* Again more cross-attention.

* More changes but I actually don't remember what.

* Hook the cross attention.

* Bugfixes.

* Another fix.

* Add the sin embeddings.

* Another fix.

* More bugfixes.

* Finally some proper generations?

* Remove some debugging.

* Hopefully a last bugfix?

* Cosmetic tweaks.

* Cross-attn caching.

* Bump the mlx version.
 2025-07-02 08:29:11 +02:00
Alexandre Défossez
e826a7034a
TTS in Moshi (#310) 
* plop

* plop

* fixes

* bump version

* nit

* plop

* update script

* adding voice script

* plop

* nit

* nit

* nit

* plop

* plop

* wip

* fixing prefix

* fix bug

* fixes

* comments

* plop

* remove unused
 2025-06-30 14:31:01 +02:00
Laurent Mazare
178b36296e
Fixes for clippy 1.88. (#311) 2025-06-30 09:39:09 +02:00
eugene-kharitonov
9a16d6c769
Bumping hf-hub dependency (#307) 
* Bumping hf-hub dependency

A version above 0.29 is needed for the OpenASR leaderboard evals.

* touching pyproject.toml too

* hf hub dep vesion >= 0.24, < 0.34 instead

---------

Co-authored-by: Eugene <eugene@kyutai.org>
 2025-06-26 12:16:57 +02:00
Laurent Mazare
e5aecfe4a2
Update the hf-hub dependency. (#306) 2025-06-23 16:34:57 +02:00
eugene-kharitonov
4f7ce70e22
removed pytorch stt inference scripts (#305) 
Co-authored-by: Eugene <eugene@kyutai.org>
 2025-06-20 17:06:14 +02:00
Laurent Mazare
1290f56dcd
Support json configs in the MLX quantization script. (#304) 2025-06-20 16:10:47 +02:00
eugene-kharitonov
26e94a6159
Reference script for STT with timestamps (#302) 
* example script

* Example script dumping timestamps

---------

Co-authored-by: Eugene <eugene@kyutai.org>
 2025-06-20 08:50:45 +02:00
eugene-kharitonov
0395bd6c9a
few last pieces in comments (#301) 
Co-authored-by: Eugene <eugene@kyutai.org>
 2025-06-19 12:45:21 +02:00
Laurent Mazare
d572ee642f
Handle the STT padding in the inference script. (#300) 
* Handle the STT padding in the inference script.

* Fix.
 2025-06-19 11:25:28 +02:00
Laurent Mazare
14012638cc
Proper audio padding for mlx. (#299) 2025-06-19 10:49:20 +02:00
eugene-kharitonov
6bef0db76b
Grabbing delay/prefix duration from json config (#298) 
Also renaming ASR->STT for uniformity

Co-authored-by: Eugene <eugene@kyutai.org>
 2025-06-19 10:28:20 +02:00
Alexandre Défossez
37bf33ace9 version v0.2.5 release 2025-06-18 09:51:46 +02:00
Laurent Mazare
2f53cc7eb4
Remove the legacy TTS. (#297) 2025-06-17 16:28:15 +02:00
Laurent Mazare
9c62b6d6bb
Bump the package version + append zeros. (#296) 2025-06-17 07:53:04 +02:00
Laurent Mazare
2a867e2fe0
Display the temperatures when using run_inference. (#295) 
* Display the temperatures when using run_inference.

* Append some silence.
 2025-06-17 07:33:53 +02:00
Laurent
2dfb291d9a Add a temperature option. 2025-06-16 17:41:55 +02:00
Laurent Mazare
c463899b46
Support inference without a depformer on the mlx side. (#294) 2025-06-16 17:37:39 +02:00
Laurent Mazare
16d2fb9ccd
Support inference without a depformer (#293) 
* Support the ASR in moshi.run_inference.

* Only assert on the absence of tokens when there is a delay.
 2025-06-16 16:40:03 +02:00
eugene-kharitonov
5a54044f31
Streaming STT example inference script (#292) 
* Streaming STT example inference script

* Streaming inference on a predefined set of HF datasets
* Includes OpenASR datasets and distil-whisper
	long-form datasets

* formatting & types

---------

Co-authored-by: Eugene <eugene@kyutai.org>
 2025-06-16 13:54:32 +02:00
Laurent Mazare
85062f3c05
Rust update. (#291) 2025-06-16 07:42:49 +02:00
Laurent Mazare
c548eedff9
Update the inference server. (#287) 2025-06-10 08:20:20 +02:00
Alexandre Défossez
3552e21871
Fix moshi run_inference with the new import_torch (#286) 
* update script

* fixing bug preventing to evaluate the old moshi models.

* fix
 2025-06-09 17:36:31 +02:00
Laurent Mazare
aa9d75d785
Embed the config files in the moshi-server binary. (#285) 2025-06-05 10:06:16 +02:00